feat(eval): pluggable benchmark harness with in-house coding-agent corpus (#562)
* feat(eval): pluggable benchmark harness with in-house coding-agent corpus
Adds eval/ tree (outside files field so npm tarball stays thin) with Adapter
interface, three reference adapters (grep / vector / agentmemory-hybrid),
two benchmarks (LongMemEval _s public, coding-agent-life-v1 in-house 15
sessions), scoring (P@K, R@K, hit, top-gold-rank), NDJSON output,
sandbox script.
coding-agent-life-v1 published scorecard at
docs/benchmarks/2026-05-20-coding-agent-life-v1.md:
agentmemory-hybrid R@5=0.967 P@5=0.578 (100% hit) vs grep R@5=0.967 P@5=0.267.
2.2x better precision on identical input, sandbox-reproducible.
Adapter contract: init(sessions, config) -> State; query(q, state, k) -> RankedDoc[]
npm scripts:
npm run eval:coding-life (no download, no API key for grep)
npm run eval:longmemeval (needs OPENAI key + 278MB download)
eval/scripts/sandbox.sh boots clean agentmemory + iii-engine on ports
3411/3412 with isolated data dir; tears down on exit.
README headline updated. 1072/1072 tests pass + 5 new eval tests.
* fix(eval): address review findings on benchmark harness
- agentmemory adapter: prefer row.sessionId before observationToSession lookup
- vector adapter: validate embedBatch response (length, indexes, non-empty rows)
- coding-life: positive-int guard on --k; wrap query loop in try/finally so teardown runs
- longmemeval: positive-int guards on --k/--limit/--stratify; per-question try/finally
- load: throw on haystack_session_ids vs haystack_sessions length mismatch
- score: P@K denominator is k (requested cutoff) not topK.length
- sandbox.sh: guard rm -rf with non-empty + /tmp/ prefix check
- README: drop unsafe rm "$(which iii)"; instruct ~/.local/bin + PATH instead; add language tag to repo-layout fenced block
- sessions.json: fix "two-phase" -> "three-phase" wording mismatch
Rohit Ghumare
7fb72f4010